Accelerate Your AI/ML Workloads With Topology-Aware Scheduling in Kueue
Talkat Cloud Native + Kubernetes AI Day Europe 2025 · on Tuesday, 2025-04-01· at 13:30· Level 1 | Hall Entrance N10 | Room G·
summary
Kueue adds topology-aware scheduling by consuming a simple node-label topology API linked to resource flavors, reconstructing a topology tree, running two BFS-like placement passes, and assigning pods with rank-based pod ordering and scheduling gates so pods start with the intended locality. Benchmarks run GPT-2 across 32 GPU nodes show about a 15% speedup in training time. Implementation requires Kueue as the job-level scheduler, supports many job types, runs in scheduling cycles with snapshots, and ships an auto mode in v0.11.
AI-written summary · about
chapters (9)
- 0:00Introduction to Kueue
- 2:11Challenges without TAS
- 4:54Motivation for TAS
- 7:32Configuring topology
- 9:03Topology trees and capacity
- 11:16User-facing annotations
- 12:48TAS mechanics and internals
- 14:44Performance and summary
- 18:05Q&A session
related talks
30mTopology-Aware Scheduling for AI Training & Inference With Kueue
28mMultitenancy and Fairness at Scale with Kueue: A Case Study
20mBoosting Training and Inference Performance via Topology-Aware Scheduling of Heterogeneous Resources
10mLightning Talk: Mind the Topology: Smarter Scheduling for AI Workloads on Kubernetes
35mKueue: A Kubernetes-native Job Queueing
33mBatch Systems in Production with Kueue: Multi-Tenancy and Fungibility
29mBuilding a Batch System for the Cloud with Kueue
38mBeyond Kubernetes: Adapting To Spe...